Interview, Fireside Chat
Inference Will Be $10TRN Market: What No One Sees | Software Investors are F***ed | Naveen Rao
- Energy supply is expected to deplete within three to five years as excess grid capacity exhausts, creating scarcity that prevents the construction of 40 gigawatts annually and drives data center power costs 30% to 40% above standard grid rates.
- The hardware sector is projected as the primary investment focus over the next decade, contrasting with "pure software" investors who face margin compression and a diminished market opportunity over a 20-year horizon.
- The inference market is forecast to reach a $10 trillion valuation within 10 years, fueled by infinite demand for energy-to-intelligence conversion and a shift where 80% of token traffic flows through open models while 80% of margins concentrate on frontier closed models within three years.
- Physical efficiency gains from traditional scaling laws have largely halted, necessitating a transition to unwaveringly specialized chips and model-to-architecture co-design, with an eventual move toward a "post-transformer world" as current attention mechanisms prove inefficient for reasoning.
- Future development bottlenecks will shift from model production to the verification of AI-generated artifacts, while AI-driven agents are expected to cover more surface area per unit of time, potentially reducing tasks from two weeks to one hour within the next couple of years.
- Robotics is anticipated to produce working prototypes in two to three years and emerge as a significant market in five to seven years, eventually expanding to become 100 times larger than the agentic coding market due to the dynamic nature of physical problem-solving.
- Investment risks include the potential for AI decision chains to fail and produce robust-looking but incorrect results, a lack of trust in agent decisions compared to human judgment, and the expectation that "customer support AI" and IDE plugins will yield low value due to non-differentiation and frictionless delivery models.
- Long-term timelines project that humans will trust AI agents more than themselves within five years, though this remains a controversial concept, while the current debt cycle is viewed as a temporary disruption that will not hinder sustained AI service demand.
- The industry will likely continue relying on pre-training and reinforcement learning for the next several years before adopting new paradigms, with small firms like "mom and pop" law firms unable to build custom models and instead relying on outsourced solutions due to scale constraints.